A Comparative Analysis Of CNN Models For Object Detection In Thermal Imaging

Abstract

Object detection in thermal imaging is a prevalent method within the fields of surveillance, medicine, and health. Thermal images represent the thermal reflections emitted by objects or bodies, captured using long- wavelength infrared (LWIR) cameras or thermal cameras. This paper reviews and compares four convolutional neural network (CNN) models—Faster R-CNN (FRCNN), Efficient Det, Single Shot Detector (SSD), and You Only Look Once (YOLO)—for object detection in thermal images. FRCNN demonstrates high accuracy, particularly for complex objects and small-scale features, but it isslower due to its two-stage process. Efficient Det is computationally efficient while maintaining high accuracy. SSD predicts object classes and bounding boxes in a single step without generatingregionproposals, whereasYOLO is highly efficient and suitable for real-time object detection. Precision, recall, and F1 score performance parameters are employed for comparison. It is observed that among EfficientDet, SSD, and FRCNN, the YOLOv8 method provides superior precision at 70.6%, recall at 43.6%, and F1 score at 53.9% for the given thermal images. These findings assist in selecting the appropriate model for object detection in thermal imaging with enhanced performance.